The vast majority of current mainstream AI programming assistants use cloud-based large model architectures, where user input is sent to the service provider's servers for processing.